Intrinsic plasticity coding improved spiking actor network for reinforcement learning.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep reinforcement learning (DRL) utilizes deep neural networks (DNNs) for significant advancements.
• Spiking neural networks (SNNs) mimic biological brain efficiency with binary signals and plasticity.
• Effective information encoding is crucial for SNNs' computational mechanisms.

Purpose of the Study:

• To develop an improved spiking actor network (IP-SAN) for reinforcement learning (RL).
• To enhance spatiotemporal state representation and biological simulation accuracy in RL agents.
• To achieve more effective decision-making in continuous control tasks.

Main Methods:

• Integration of adaptive population coding at the network level.
• Incorporation of dynamic spiking neuron coding at the neuron level.
• Development of an intrinsic plasticity coding mechanism for SNNs.

Main Results:

• The proposed IP-SAN demonstrated superior performance compared to state-of-the-art methods.
• The model achieved significant improvements in five continuous control tasks.
• Enhanced spatiotemporal state representation and biological simulation accuracy were observed.

Conclusions:

• The IP-SAN offers a promising approach for efficient decision-making in RL.
• The integration of intrinsic plasticity and advanced coding strategies enhances SNN performance.
• This work contributes to more biologically plausible and effective AI agents.